Multi-fidelity reinforcement learning framework for shape optimization